Solving Concurrent Multiagent Planning using Classical Planning - - PowerPoint PPT Presentation

▶
solving concurrent multiagent planning using classical
SMART_READER_LITE
LIVE PREVIEW

Solving Concurrent Multiagent Planning using Classical Planning - - PowerPoint PPT Presentation

Jun 20, 2023 382 likes •610 views

Introduction Background Compilation Experiments Conclusions Solving Concurrent Multiagent Planning using Classical Planning Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra June 26, 2018 Daniel Furelos-Blanco and Anders

slide-1
SLIDE 1

Introduction Background Compilation Experiments Conclusions

Solving Concurrent Multiagent Planning using Classical Planning

Daniel Furelos-Blanco and Anders Jonsson

Universitat Pompeu Fabra

June 26, 2018

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-2
SLIDE 2

Introduction Background Compilation Experiments Conclusions

Motivation

Lot of progress in multiagent planning since CoDMAP-15. Limitation: benchmark domains can be solved using sequential plans. Many applications require agents to act in parallel, like RoboCup Soccer or RoboCup Rescue.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-3
SLIDE 3

Introduction Background Compilation Experiments Conclusions

Proposed approach

Solve multiagent planning problems that involve concurrency by translating them into classical planning. Concurrency expressed using concurrency constraints which model when

1 two actions must occur in parallel, or 2 two actions cannot occur in parallel.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-4
SLIDE 4

Introduction Background Compilation Experiments Conclusions

Concurrent Multiagent Planning - Definition

A classical planning problem is defined as Π = F, A, I, G where

F is a set of fluents, A is a set of atomic actions, I ⊆ F is an initial state, and G ⊆ F is a goal condition.

A concurrent multiagent planning problem (CMAP) is a tuple Π =

  • N, F,
  • Ain

i=1 , I, G

  • where N = {1, . . . , n} is the agent set, and Ai is the action

set of agent i ∈ N.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-5
SLIDE 5

Introduction Background Compilation Experiments Conclusions

Concurrent Multiagent Planning - Joint Actions

Each action is a joint/concurrent action: a combination of atomic actions simultaneously performed. Given a concurrent action a =

  • a1, . . . , ak

, its precondition and effects are defined as pre(a) =

k

  • j=1

pre(aj), eff(s, a) =

k

  • j=1

eff(s, aj) Constraints are imposed on atomic actions to ensure joint actions are well-defined.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-6
SLIDE 6

Introduction Background Compilation Experiments Conclusions

Concurrent Multiagent Planning - Concurrency Constraints

Formulation in [Boutilier and Brafman, 2001] (later extended in [Kovacs, 2012]) uses actions as fluents:

Positive: action a1 has a2 as precondition. Negative: action a1 has ¬a2 as precondition.

Effects of an action a1 can be conditioned to the simultaneous execution of another action a2. Implicit negative concurrency constraint:

Each agent contributes at most once to the joint action.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-7
SLIDE 7

Introduction Background Compilation Experiments Conclusions

Concurrent Multiagent Planning - Example

TableMover [Boutilier and Brafman, 2001]: Two agents must move blocks between rooms. Put blocks on a table, carry the table together to another room, and tip the table to make the blocks fall down.

a1 a2 b1 r1 r2 s1 s2 Table

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-8
SLIDE 8

Introduction Background Compilation Experiments Conclusions

Concurrent Multiagent Planning - Example

(:action lift-side :agent ?a - agent :parameters (?s - side) :precondition (and (at-side ?a ?s) (down ?s) (handempty ?a) (forall (?a2 - agent ?s2 - side) (not(lower-side ?a2 ?s2)) ) ) :effect (and (not (down ?s)) (up ?s) (lifting ?a ?s) (not (handempty ?a ?s)) ... ... (forall (?b - block ?r - room ?s2 - side) (when (and (inroom Table ?r) (on-table ?b) (down ?s2) (forall (?a2 - agent) (not (lift-side ?a2 ?s2)) ) ) (and (on-floor ?b) (inroom ?b ?r) (not (on-table ?b)) ) ) )))

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-9
SLIDE 9

Introduction Background Compilation Experiments Conclusions

Compilation from Multiagent to Classical Planning (I)

Transform a CMAP Π =

  • N, F,
  • Ain

i=1 , I, G

  • into a

classical planning problem Π′ = F ′, A′, I′, G′. Sound and complete transformation:

Adds new fluents and actions that allow to select and apply joint actions while respecting concurrency constraints.

Divide simulation of a joint action in three different phases:

1 Action selection: check preconditions of constituent atomic

actions.

2 Action application: apply effects of constituent atomic

actions.

3 Resetting: reset auxiliary fluents.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-10
SLIDE 10

Introduction Background Compilation Experiments Conclusions

Compilation from Multiagent to Classical Planning (II)

select-phase select-ai apply-phase do-ai reset-phase end-ai finish repeat t, 1 ≤ t ≤ n repeat t repeat t Joint action Start new joint action

The resulting number of actions is polynomial, not exponential:

  • A′

= 3

  • i∈N
  • Ai

+ 4.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-11
SLIDE 11

Introduction Background Compilation Experiments Conclusions

Compilation from Multiagent to Classical Planning (III)

Extension: joint actions with bounded size C. At most C agents can act at a time. Purpose: reduce branching factor. The number of actions is still polynomial:

  • A′

= (2C + 1)

  • i∈N
  • Ai

+ 4.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-12
SLIDE 12

Introduction Background Compilation Experiments Conclusions

Compilation from Multiagent to Classical Planning (IV)

Example

a1 a2 b1 r1 r2 s1 s2 Table

Multiagent plan

1 (to-table a1 r1 s2)(pickup-floor a2 b1 r1) 2 (putdown-table a2 b1 r1) 3 (to-table a2 r1 s1) 4 (lift-side a1 s2)(lift-side a2 s1) 5 (move-table a1 r1 r2 s2)(move-table a2 r1 r2 s1) 6 (lower-side a1 s2)

Classical plan (1st joint action)

1 (select-phase ) 2 (select-to-table a1 r1 s2) 3 (select-pickup-floor a2 b1 r1) 4 (apply-phase ) 5 (do-pickup-floor a2 b1 r1) 6 (do-to-table a1 r1 s2) 7 (reset-phase ) 8 (end-to-table a1 r1 s2) 9 (end-pickup-floor a2 b1 r1) 10 (finish )

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-13
SLIDE 13

Introduction Background Compilation Experiments Conclusions

Experiments

Tests on two sets of domains:

1 CoDMAP-15 domains (do not require concurrency). 2 Domains that require concurrency:

TableMover [Boutilier and Brafman, 2001]. Maze [Crosby et al., 2014]. BoxPushing [Brafman and Zoran, 2014]. Workshop.

Test three variants of the compilation + Fast-Downward: Unbounded (∞). Joint action size ≤ 2 (C = 2). Joint action size ≤ 4 (C = 4).

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-14
SLIDE 14

Introduction Background Compilation Experiments Conclusions

Experiments - CoDMAP-15 domains (I)

Used to show that our algorithm is complete. However, they do not require concurrency (can be sequentially solved).

Use Fast-Downward (FD) for comparison.

They do not specify negative concurrency constraints (incompatible actions).

Add explicit negative concurrency constraints to avoid invalid plans → Complexity to find a plan increases!

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-15
SLIDE 15

Introduction Background Compilation Experiments Conclusions

Experiments - CoDMAP-15 domains (II)

Domain # Coverage Time (s.) Plan length # Actions 2 4 ∞ FD 2 4 ∞ FD 2 4 ∞ FD 2 4 ∞ FD Blocksworld 20 7 2 4 20 759.5

  • 0.2

32.1

  • 32.8

6848 12323 4110 1270 Depot 20 13 10 9 17 202.9 246.4 223.9 58.3 30.6 15.7 14.9 44.0 10100 18176 6061 2007 Driverlog 20 18 17 18 20 67.3 58.8 73.7 26.1 21.1 20.5 25.2 35.6 38416 69145 23051 7386 Elevators08 20 9 8 10 20 13.8 12.5 9.5 0.2 31.0 30.3 36.4 65.1 10779 19399 6469 2155 Logistics00 20 20 20 20 20 1.9 2.9 212.1 0.0 30.3 28.0 30.1 50.2 1781 3202 1070 318 Rovers 20 20 20 19 20 45.2 75.2 20.9 0.1 46.5 47.4 42.9 56.8 18314 32962 10990 2609 Satellites 20 19 17 19 20 82.8 128.0 32.2 1.0 32.6 35.5 34.2 55.9 45106 81188 27065 8122 Sokoban 20 18

  • 32.3
  • 54.1

3319 5970 1993 663 Taxi 20 20 20 20 20 1.3 2.6 0.7 0.0 14.8 14.7 14.7 18.7 544 975 328 108 Wireless 20 2 2 2 4

  • 15644

28156 9388 3128 Woodworking08 20 14 8 4 20 290.0 256.0

  • 0.9

22.4 11.4

  • 46.1

17406 31327 10445 3447 Zenotravel 20 16 16 18 20 87.2 125.6 164.8 1.5 23.6 24.1 34.2 46.9 67586 121652 40553 13502 Total 240 158 140 143 219

Lower coverage and slower than FD. Solutions are always shorter (FD does not compress plans).

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-16
SLIDE 16

Introduction Background Compilation Experiments Conclusions

Experiments - Required Concurrency Domains (I)

TableMover - Move blocks between rooms using a table.

The table must be moved simultaneously. The blocks on the table fall if only one side is lifted.

Maze - Move between two cells in a grid using:

Doors: traversed only by one agent at a time. Bridges: can be traversed by multiple agents at once. Boat: used by two or more agents at once (same direction).

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-17
SLIDE 17

Introduction Background Compilation Experiments Conclusions

Experiments - Required Concurrency Domains (II)

BoxPushing - Push boxes between two locations in a grid.

A small box requires 1 agent to push. A medium box requires 2 agents to push. A large box requires 3 agents to push.

Workshop - Inventory pallets in a high-security facility.

Open door = press switch + turn key. Inventory a pallet = lift pallet + examine pallet.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-18
SLIDE 18

Introduction Background Compilation Experiments Conclusions

Experiments - Required Concurrency Domains (III)

Compare our approach with CJR [Crosby et al., 2014]: Compilation to classical planning. Concurrency constraints in the form of affordances on subsets

  • f objects.

Limitations:

Concurrency constraints are not as expressive → Conditional effects on simultaneous actions are not supported. Effects are applied immediately for atomic actions → Some joint actions cannot be simulated.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-19
SLIDE 19

Introduction Background Compilation Experiments Conclusions

Experiments - Required Concurrency Domains (IV)

Domain # Coverage Time (s.) Plan length # Actions 2 4 ∞ CJR 2 4 ∞ CJR 2 4 ∞ CJR 2 4 ∞ CJR Maze 20 13 8 6 11 351.9 435.2 144.4 192.8 47.2 22.0 11.7 77.3 41723 69368 27900 156886 a = 10 10 8 6 5 7 243.6 564.8 169.1 225.5 48.3 25.0 12.2 79.6 39909 67417 26155 119374 a = 15 10 5 2 1 4 525.2

  • 45.4
  • 43989

71807 30080 194397 TableMover 24 15 12 15

  • 263.3

336.5 341.0

  • 58.7

59.0 61.5

  • 7487

13127 4667

  • a = 2

12 10 10 11

  • 103.8

226.4 214.6

  • 63.5

62.0 64.5

  • 3450

6154 2098

  • a = 4

12 5 2 4

  • 558.2
  • 49.0
  • 11524

20100 7236

  • Workshop

20 15 13 13 6 132.3 298.6 51.8 629.0 35.7 37.0 32.5 63.5 18002 31000 11502 5425 a = 4 10 8 8 8 5 42.1 261.6 36.6 587.3 37.3 43.9 37.3 65.8 7772 13621 4847 2351 a = 8 10 7 5 5 1 235.5 357.8 76.0

  • 33.9

26.0 24.8

  • 28231

48378 18157 8499 BoxPushing 69 39 56 59

  • 26.8

79.9 63.9

  • 9.4

11.0 10.2

  • 3075

5360 1932

  • a = 2

21 21 21 21

  • 14.1

15.6 16.2

  • 10.5

10.6 10.3

  • 2099

3775 1261

  • a = 3

48 18 35 38

  • 41.5

118.5 90.2

  • 8.2

11.2 10.2

  • 3502

6054 2226

  • l = 0

21 18 17 18

  • 41.5

64.7 53.3

  • 8.2

8.3 8.3

  • 3373

5887 2116

  • l > 0

27

  • 18

20

  • 169.2

123.5

  • 13.9

12.0

  • 3602

6184 2312

  • Unbounded compilation (∞) has the highest coverage.

Compilation C = 2 is usually fast but cannot solve problems involving > 2 agents.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-20
SLIDE 20

Introduction Background Compilation Experiments Conclusions

Conclusions

Sound and complete method for compiling CMAPs into classical planning problems. The number of resulting actions is polynomial in the description of the CMAP. Competitive performance in CoDMAP-15 domains and domains requiring concurrency.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-21
SLIDE 21

Introduction Background Compilation Experiments Conclusions

Questions

Contact:

daniel.furelos@upf.edu anders.jonsson@upf.edu

Software: https://github.com/aig-upf/ universal-pddl-parser-multiagent

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning

slide-22
SLIDE 22

Introduction Background Compilation Experiments Conclusions

Boutilier, C. and Brafman, R. I. (2001). Partial-Order Planning with Concurrent Interacting Actions.

  • J. Artif. Intell. Res. (JAIR), 14:105–136.

Brafman, R. I. and Zoran, U. (2014). Distributed Heuristic Forward Search with Interacting Actions. In Proceedings of the 2nd ICAPS Distributed and Multi-Agent Planning workshop (ICAPS DMAP-2014). Crosby, M., Jonsson, A., and Rovatsos, M. (2014). A Single-Agent Approach to Multiagent Planning. In ECAI 2014 - 21st European Conference on Artificial Intelligence, 18-22 August 2014, Prague, Czech Republic - Including Prestigious Applications

  • f Intelligent Systems (PAIS 2014), pages 237–242.

Kovacs, D. L. (2012). A Multi-Agent Extension of PDDL3.1. In Proceedings of the 3rd Workshop on the International Planning Competition (IPC), pages 19–27.

Daniel Furelos-Blanco and Anders Jonsson Universitat Pompeu Fabra Solving Concurrent Multiagent Planning using Classical Planning